CA2252428A1 - Nonlinear-approximator-based automatic tuner - Google Patents
Nonlinear-approximator-based automatic tuner Download PDFInfo
- Publication number
- CA2252428A1 CA2252428A1 CA002252428A CA2252428A CA2252428A1 CA 2252428 A1 CA2252428 A1 CA 2252428A1 CA 002252428 A CA002252428 A CA 002252428A CA 2252428 A CA2252428 A CA 2252428A CA 2252428 A1 CA2252428 A1 CA 2252428A1
- Authority
- CA
- Canada
- Prior art keywords
- parameters
- approximator
- nonlinear
- tuner
- controller
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Abandoned
Links
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
- G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
- G05B13/0265—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion
- G05B13/027—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion using neural networks only
Abstract
An automatic tuner for control systems that produces as output values for parameters of an arbitrary controller. The controller is in a control loop so that its output effects changes in actuators and regulates a physical process.
The controller consists of either linear and nonlinear controller components or a combination of both. The tuner has a nonlinear approximator that has been optimized off-line. The off-line optimization is done without supervised learning so that desired outputs of neural network do not need to be available, and separate optimization to generate the desired outputs is not necessary. The off-line optimization can also rely on optimization criteria that are arbitrary: differentiability, convexity, even continuity of criteria are not required. The off-line optimization ensures robustness of generated controller parameters so that the input process characteristics do not need to be highly accurate. The off-line optimization is performed in such a way that the input parameters that relate to desired closed-loop system behavior include robustness parameters that can be used to effect tradeoffs between robust and nominal performance. The inputs to the nonlinear approximator consist of two sets of input parameters, either of which may be empty. A first set of input parameters can include parameters that relate to process characteristics. A second set of input parameters can include parameters that relate to desired closed-loop system behavior. The output values may be proportional and/or integral and/or derivative gains for PID-like controllers.
The output values otherwise may be parameters for delay-compensation controllers, parameters for controllers that consist of lead-lag terms in combination with PID controllers, parameters for higher-order linear controllers, discrete variables that select between different control structures, or parameters for nonlinear controllers of predetermined structure. The nonlinear approximator may be implemented as a compositional sigmoidal mapping, a multilayer perception structure, a fuzzy logic model, a radial basis function network, a polynomial expansion, or other parametrized nonlinear structure.
The controller consists of either linear and nonlinear controller components or a combination of both. The tuner has a nonlinear approximator that has been optimized off-line. The off-line optimization is done without supervised learning so that desired outputs of neural network do not need to be available, and separate optimization to generate the desired outputs is not necessary. The off-line optimization can also rely on optimization criteria that are arbitrary: differentiability, convexity, even continuity of criteria are not required. The off-line optimization ensures robustness of generated controller parameters so that the input process characteristics do not need to be highly accurate. The off-line optimization is performed in such a way that the input parameters that relate to desired closed-loop system behavior include robustness parameters that can be used to effect tradeoffs between robust and nominal performance. The inputs to the nonlinear approximator consist of two sets of input parameters, either of which may be empty. A first set of input parameters can include parameters that relate to process characteristics. A second set of input parameters can include parameters that relate to desired closed-loop system behavior. The output values may be proportional and/or integral and/or derivative gains for PID-like controllers.
The output values otherwise may be parameters for delay-compensation controllers, parameters for controllers that consist of lead-lag terms in combination with PID controllers, parameters for higher-order linear controllers, discrete variables that select between different control structures, or parameters for nonlinear controllers of predetermined structure. The nonlinear approximator may be implemented as a compositional sigmoidal mapping, a multilayer perception structure, a fuzzy logic model, a radial basis function network, a polynomial expansion, or other parametrized nonlinear structure.
Description
CA 022~2428 1998-10-20 ~ Z / G~6/~
NONLINEAR-~PPRO,YIMATOR-BASED AUTOMATIC TUNER
BACKGROUND OF THE INVENTION
The present invention pertains to tuners for controllers, and particularly to automatic tuners for controllers. More particularly, the invention pertains to nonlinear-S approximator-type of tuners. U.S. Patent No. 5,311,421, by Masahide Nomura et al., issued May 10, 1994, and entitled "Process Control Method and System for Performing Control of a Controlled System by Use of a Neural Network," provides background information for the present invention and is hereby incorporated by reference in this description. Reference may also be had to German Patent DE 4433332, which describes the use of a neural network as a tuner for a controller.
SUMMARY OF THE INVENTION
The invention is a technique for developing a tuner which is used for tuning or optimally guiding a controller. The tuner has a preprocessor for transforming a set of input signals into a set of normalized parameters. These parameters are inputted to a nonlinear approximator that has been setup off-line without advance supervised learning using a plurality of closed-loop simulations. An optimization algorithm minimi7es the value of a cost function which is derived from the closed-loop simulations and which indicates the performance of the simulations, to produce optimized parameters of the nonlinear approximator. The nonlinear approximator outputs a set of normalized tuning parameters, derived on the basis of the minimi7ed value of the cost function, to a postprocessor which scales this set of parameters into controller tuning parameters which go to the controller. The nonlinear approximator can be based on neural networks or other parametrized nonlinear structures. In essence, deterrnining nonlinear approximator parameters actually amounts to designing or setting-up a tuner. In turn, the tuner tunes the controller having a closed loop, with appropriate controllerparameters.
BRIEF DESCRIPTION OF THE DRAWING
Figure 1 shows a basic controller and process relationship.
Figure 2 shows a controller with a tuner, along with a process.
Figure 3 shows a process with a step-function input.
Figure 4 shows the characteristics of an open loop step response of the process in figure 3.
~ENDED
., . .. -- ................ . . ....... . ... ..
CA 022~2428 1998-10-20 22/C~
Figure 5 shows a graph with a step tunction having a resultant overshoot and an oscillatory settling which has to be modeled as at least a second order system with delay.
Figure 6 reveals the closed-loop step response parameters of a system.
,~E~E~
. _ ~ . .
CA 022~2428 1998-10-20 W O 98/00763 PCTrUS97/09473 Figure 7 reveals a tuner with separate inputs for process characteristic parameters and closed-loop performance parameters.
Figure 8 is a graph of fast settling performance with overshoot.
Figure 9 is a graph of slow settling performance without overshoot.
S Figures l Oa and I Ob show an automatic tuner for processes that can be modeled as first-order linear systems with delay.
Figure 11 shows a neural network tuner used in conjunction with a parametrized neurocontroller.
Figure 12 illustrates a prior art neural network tuner with a supervised learning l O algorithm feature.
Figure 13 is a schematic of one framework of the present invention, not relying on supervised learning.
Figure 1 4a shows an example of a tuner for a proportional-integral (PI) controller embodied as a neural network with one tuning knob input and outputs for proportional and integral gains.
Figure 1 4b shows a tuner embodied as a computationally simple compositional mapping with one tuning.
Figure 15 shows how the tuner of figure 14 can be used for a known linear first-order process.
Figure 16 is a graph for slow or fast settling, dependent upon tuner adjustment input.
Figure 17 is a sçh~n ~tic for developing the tuner of figure 15 using a nonlinear optimization algorithm.
DESCRIPTION OF THE EMBODIMENT
Figure 1 shows a simple system proportional-integral-derivative (PID) controller12 that may be used for control of the t~ dlllre in a room. The setpoint Yd for temperature is input at line 13 to controller l 4 which has a u output on line 15 for controlling an airflow, such as heat or hot air flow, to process 16. The controller output goes to the furnace via a gas valve or other furnace controlling device. Process 16 is the room having its space le~ ld~ controlled. An output y on line 18 indicates the ten~ re of the room. Signal y is fed back to an input 20 of controller 14. In other words, controller 14 looks at the current temperature on line 18 and the setpoint CA 022~2428 1998-10-20 W O 98/00763 PCTrUS97/09473 (thermostat) temperature (y desired) on line 13, and uses the difference between the temperatures to calculate heating input 15 to the room or process 16.
"u" can be shown relative to various domains in a simple proportional-integral (PI) controller according to the following equations, where:
kc is the proportional gain for the controller;
k; is the integral gain for the controller;
t is time;
~ is an integration variable;
e(t) is the process output error at time t;
u(t) is the controller output at time t;
y(t) is the process output at time t;
yd(t) is the desired process output or setpoint of time t;
U(s) is the Laplace transform of u(t);
E(s) is the Laplace transform of e(t);
j is a summ:-tion index; and I is a discrete time index.
u(t) = kce(t) + k j Je(r)dr ( I ) where e(t) = Yd (t) - y(t) U(s) = kcE(s) + ' E(s) (2) u(l) = kce(l) + k, ~e(j)~t (3) j=o Equation (1) is stated for the continuous-time domain, equation (2) for the Laplace or frequency domain, and equation (3) for the discrete-time domain. The proportional (kc) and integral (kj) gains are set to values a~plopl;ate for the application 25 of control system 14 to process 16. ~t is the sampling interval.
CA 022~2428 1998-10-20 W 098/00763 PCTrUS97/09473 --4--The algorithms, used for setting controller gains, for instance kc and ki, are called "tuning algorithrns." The tuning algorithm is implemented in a tuner 21 of figure 2. The tuner puts gains kc and kj into controller 14 via line 22. To set these gains is often done by a "trial and error" approach in the related art. Current tuners in the art are 5 not wholly satisfactory.
Inputs to tuner 21 need to indicate two types of inforrnation. The first type ofinformation is the relevant dynamics of the process (e.g., characteristics of the room, such as its heat loss~ heating-up time constant, and the delay due to ductwork from the furnace to the room). The second type of information is desired dynarnics of the closed 10 loop (e.g., time to reach the setpoint, settling time, overshoot, tolerances of overshoot, and importance of reducing control energy).
Figure 3 shows a step function 24 of u to input 15 of process 16 and output y from line 18. Figure 4 shows the characteristics of an open loop step response of process 16. The process is of a first order with a delay linear system which is usually 15 sufficient for most industrial and building control processes. Such system has three process model parameters Kp, Tp and TD.
Thegainis K = (y,- yo)/(u,- 1~o) The time constant, T = t2 ~ t" (which is the time that y takes to reach 63 percent of the final output). The dead time or delay is T" = t,--to . If Kp, Tp, TD~ are changed, 20 then the curve for y will change.
Figure 5 shows a graph with a step function having a resultant overshoot and an oscillatory settling 25 which has to be modeled as at least a second order system with delay, relative to input 24.
Figure 1 shows a closed-loop system 12 with a step function signal 24 to input 13. Figure 6 reveals the closed-loop step response parameters of system 12.
Process overshoot is y~5 =
Settling time is t,.7 = t, - to Rise time is tpy~ = t2--to Control energy is Ec - ~(u(r) - u(r- 1))2 CA 022~2428 1998-10-20 W O 98/00763 PCTfUS97/09473 One important objective in tuning a controller is to manipulate these and other closed-loop response parameters. For example, large overshoots are typically undesirable~ but often unavoidable if fast settling times are required (the settling time is the time after the setpoint change after which the process output y is within some S quantity, e.g. 95%, of the desired output or setpoint Yd, with the percentage computed relative to the magnitude of the setpoint change). For fast settling, control energies will also typically be high. The response parameters have a complex relationship to each other and to the tuning parameters, and no formulae are available that can accurately characterize this relationship.
Different tuning methods take different types of tuner inputs of the relevant dynamics of the process and desired dynamics of the loop, and provide gains for different types of controllers. The present invention of tuning can be used for any kind of linear or nonlinear controller, not just a PID type. Tuner inputs can be customized as desired, and the tuner can be made robust to whatever extent desired. Robustness is the 15 graceful tolerance of uncertainty of knowledge about the system. It is also the graceful tolerance of change or drift in the system.
Two kinds of parameters are "nominal" and "estimated". For example, Kp is the nominal or actual gain and Kp is estimated gain. Controllers are designed to take into consideration that Kp ~ Kp.
Features of the present invention are the use of one or several nonlinear approximators in the tuner, and the use of a general optimization system. The following types of nonlinear approximators are explicitly considered:
~ Conventional multilayer perceptron neural networks ~ Computationally simple, compositional sigmoidal nonlinear mappings ~ Radial basis function networks ~ Functional link networks ~ Cerebellar model articulation controller (CMAC) networks ~ Fuzzy logic models ~ Wavelet networks ~ Polynomial expressions .. .. . . . .. . . .
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 ~ Specific nonlinear parametrized structures based on human intuition about controller tuning The common aspect of these and other relevant approximator structures is that they implement nonlinear mappings the form of which can be adjusted by manipulating 5 parameters associated with the approximator. Thus generally a nonlinear approximator, as used in this invention, can be expressed as follows:
Pc = F(I; Pa) where Pc is the vector of controller tuning parameters, I is the input vector to the nonlinear approximator, and Pa is the parameter vector for the approximator. For10 example, a conventional multilayer perceptron neural network may implement the mapping F as:
- PC~ +exp -~,w,lhj+b~ ) 15 where the hi are intermediate ("hidden unit") variables computed as:
[ k = 1 D
Here the parameter vector pa consists of the input-to-hidden weights w , the hidden-20 to-output weights wij, and the hidden and output bias weights bi and b3 . i and k are summation indices, j is the element of the approximator output vector, h is the number of hidden units, and n is the number of elements of the input vector I. These expressions relate to "one-hidden-layer" neural networks. Multiple layers of hidden units can also be used.
As another example, a computationally simple, compositional sigmoidal nonlinear mapping may implement F as:
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 pC~j] ~vii l+jI.wi + ail where the parameter vector Pa comprises the vectors wi and the scalars ai and vij. i is a summation index and b is the number of nonlinear elements in the structure. The period (.) denotes the dot product operator. Note that this form of the nonlinear approximator, which otherwise resembles a neural network, does not require the use of transcendental functions such as exponentials or hyperbolic tangents which are costly to compute. This approximator is the approach of first choice for this invention. Again, multiple "layers"
of the nonlinearities can be used.
As a third exarnple, a radial basis function network may implement the mapping Fas:
i=N 2 pClj] = ~ wjj exp~lI - ,Ui ll/~ ) j=, i where the parameter vector Pa comprises the vectors ,ui and the scalars ~i and wij. i is a summation index and N is the number of Gaussian nonlinearities in the structure.As a final example, a polynomial expansion may be used as the nonlinear approximator such as:
k=n k=r~=k pC~j]= ~aklk + ~ ~bklIkIl k=1 k=ll=l where the parameter vector Pa comprises the scalars ak and bkl and n is the number of elements in the input vector I. k and 1 are s~lmm~tion indices.
Figure 7 reveals a controller system 30 which has a nonlinear-approximator-based automatic tuner 21 having separate inputs for process characteristic parameters and closed loop performance parameters. The latter can be used to adjust closed-loop performance for slow or fast settling. Inputs 26 and 27 permit one pre-designed tuner 21 to be used for many applications or be universal for large and various classes of applications. A nonlinear approximator 28 is optimized off-line but without using supervised learning methods. Nonlinear approximator 28 may be a neural network but is not used as a nonlinear regression model (as it is in Nomura et al. -- see their Fig. 33).
.. ..... . . . . . . ... . . .. .. .
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 In addition to neural network or other nonlinear approximator 28, automatic tuner 21 can contain preprocessing and post-processing functions. These include scaling functions which are done outside of the neural network. Scaling is performed mainly for linearity purposes. Process characteristic parameters 26 (notated as pp later) 5 and closed-loop performance parameters 27 (notated as pj later) allow one pre-designed automatic tuner 21 to be used for a broad variety of applications. No on-line orapplication-specific training or optimization of the neural network, or modification of other components of the automatic tuner, are needed.
A specific example is neural-network-based automatic tuner 21 for processes 10 that can be modeled as first-order with dead time linear processes. Such processes can be modeled with three process model (or process characteristic) parameters 26: the process gain, time constant, and dead time. One closed-loop performance parameter is also assumed: a settling time knob dST (this constitutes the closed loop performance parameter at 27 in figure 7), which can be used to adjust closed-loop performance for 15 slow or fast settling. There may be other closed-loop performance parameters on input 27. In this case, the neural network can be off-line optimized without using supervised learning that is used in the related art (as discussed below) using a simulation-based optimization system.
Figure 8 shows the fast settling (i.e., low tsT) performance with overshoot for a low dsT input 27. Figure 9 shows the slow settling (i.e., high tST) performance without overshoot for a high dST input 27. There may be a knob for dST that provides an analog-like variation, such as from zero to one, to automatic tuner 21.
The key feature of the present invention is the design of nonlinear approximator28. During operation in a setup similar to figure 7, automatic tuner 21 can have a basic architecture, with pre-processor, non-linear approximator 28 and post-processor, as shown in figure I Oa. Note in figure 1 Ob that three process characteristic parameters are input to automatic tuner 21: estimates of process gain, time constant, and dead time ,~ A ,~
(notated K~, T~, and T~, ). Output 22 of the tuner 21 consists of controller parameters (for example, for a PID controller, these would be the proportional, integral, and 30 derivative gains) PC In figure 10b, phias iS an internal variable to automatic tuner 21 that consists of baseline controller parameters which are then modified in a context-. . .
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 sensitive way by neural network or a nonlinear approximator 28. Note that by intelligently structuring autotuner 21, the number of inputs required for nonlinear approximator 28 is reduced -- in this case, to two. T" / T to line 46 and dST to line 29 are inputted to nonlinear approximator 28. An output 31, PC, which is a control S parameter, is fed into s~-mming junction 32. A pc bias is added at line 33 to Pc at summer 32 which outputs a pcnom, which is the vector of nominal (unscaled) control parameters, at line 34 and is inputted to scaling mechanism 35. Output 22 of scaling 35 goes to controller 14.
Tuner 21 can be used in control structure having both linear and nonlinear l 0 components. For example, tuner 21 can be used in conjunction with a parametrized neurocontroller 36, as shown in figure 11.
Neural network based automatic tuner 21 is optimized off line using a design framework. An "evolutionary computing" algorithm has been developed that incorporates aspects of genetic algorithms. Also, a gradient search can be enabled, with 15 gradients numerically computed.
The flexible nature of the design framework perrnits one to optimize for criteria that more conventional approaches seldom permit. In particular, performance criteria need not be quadratic, or even differentiable; control structures can be arbitrarily nonlinear; and robustness can explicitly be considered. The present approach to the 20 optimization of neural network based automatic tuner 21 does not require "learning in advance" in which input-output combinations must be first compiled using a separator optimization program, and neural network 28 then trained using supervised learning.
The present approach can thus be contrasted to the approach of Nomura et al. (U.S.
Patent 5,311,421) in which the use of a neural network for tuning is also disclosed but in 25 which the development of the neural network requires a supervised learning algorithm for which a~lopllate teaching data must be collected.
One may use supervised learning of Nomura et al. for neural network tuner 37 of figure 12. For tuner 37, inputs may be ~lesign~ted x and outputs designated Pc~ dST is part of x. Pc are control parameters. Pc is compared with the desired output Pc* and the 30 result of the comparison at comparator junction 38 is error ~, which is fed to learning algorithm 39. Algorithm 39 then provides an output to neural network 40 to adjust the .. .. . ..... . .... .. ...
CA 022~2428 1998-10-20 2/~ o/~
lo-weights for output adjustments. For that supervised learning, a large database of (Y, Pc*) pairs need to be compiled in advance. Pc* needs to be computed for many different cases in advance. Nomura et al. involves the inputting of inforrnation with the characteristics contained therein to a neural network 40 for "learning in advance a correlation" between the information containing the characteristics and a control parameter and determining a control parameter for a controller. "Learning in advance"
means the deterrninin~ of optimal control parameters corresponding to different characteristics of the model of the combined controller-controlled system and then using the optimal control parameters as learning teacher data. In Nomura et al., it is also assumed that input-output combinations are available as learning data (col. 11, lines 1 1-17). The Cj's in equation (10) are ideal outputs that the neural network is trained to match, and thus Cj's are to be available for neural network training. The neural network is used as a nonlinear regression model and thus the neural network is regressed using input-output training data.
l S The present approach does not rely on such supervised le~rning No collection of (x, Pc*) pairs is needed and no computation of (optimal) Pc*~s is required. The present neural network is developed using off-line optimi7~tion without any need for "learning in advance" in the sense of Nomura et al. No input-output training data is needed.
It should be noted that the prior art of Nomura et al., by virtue of its reliance on a supervised learning approach, is restricted to certain types of neural network tuners.
Our invention eschews supervised learning entirely in favor of a simulation-based optimization approach, and an important ancillary benefit of this approach is that the tuner can comprise ;~bil~ nonlinear approximators, including but certainly not limited to the type of neural network structure employed in the cited prior art.German Patent DE 4433332 also describes the use of a neural network as a tuner for a PID controller. As in Nomura et al., learning data for the neural network must first be produced before the neural network itself can be optimized.
Figure 13 shows the design framework 41 of the present invention schematically. For each parameter vector p, closed-loop simulations are run and a performance criterion value computed. No learning teacher data is required. The vector p can contain gains for a linear control structure, so that one can optimize a fixed controller for some criterion; it can contain parameters for a nonlinear-approximator-o~~s~
CA 02252428 1998-10-20 ~2/C6/~/~
-IOA
based tuner or neural network controller (e.g., pararnetrized neurocontroller (PNC) 36 in figure 11); or it can contain pararneters for both the tuner ~nd controller, so that both O~) S~
.. . .. . . . ..
CA 022~2428 1998-10-20 modules can simultaneously be optimized. The objective of the design activity is to develop tuners or controllers that can be used for a broad range of applications. Thus, numerous closed-loop simulations must be run for every choice of p. In these simulations, various parameters are varied depending on the design requirements:5 ~ Process model parameters pp which are input to the process model (PM) in the closed-loop simulation ~ Estimation errors between these process model parameters and the process model parameter estimates pr, of which the latter are input to the tuner and/or controller l O ~ Initial conditions, especially in the case that the closed-loop controller or process model is nonlinear ~ Setpoint profiles, also for nonlinear controllers or nonlinear process models ~ Parameters PJ that affect the cost function--the use of these parameters allow us to develop one control solution for a class of parametrized criteria Each closed-loop simulation results in one elementary cost function Jj being computed. The results of all the simulations conducted for one vector p are composed into an overall cost function J which is returned to the optimization algorithm:
20 J = fJ(Jl~ J2, ~ ~, JN) where N is the number of closed-loop simulations performed (typically l O00).
The above paragraphs describe the general framework of the design system. For this project, one has made specific choices for optimization runs. Note these below:
~ Process models are first order with dead time. The process gain Kp is norm~li7~1 to 1.0, the process time constant Tp to 10Ø The process dead time Td varies so that the constraint in the next statement holds:
CA 022~2428 1998-10-20 ~ Estimates of Kp, Tp, and Td are input to the neural network controller or tuner.
These estimates are perturbations of the true model parameters as discussed below. Td is limited to the range [0,2 T ].
5 ~ The parameter estimate for a parameter Pi is bounded uniformly within [( I -r)pi, (l+r)pi], where r is a parameter. r is a function ofthe criterion parameter dST
which can vary between 0 and 1. This function is a linear function so that the minimum value of r, achieved when dST is 0, is rmin~ and the maximum value, corresponding to dST equal to 1, is rmaX~ Thus, dST controls the robustness of the control, in addition to being a settling time knob as discussed below.
~ The only criterion parameter employed is dST. The elementary criterion is:
Ji - aOdSTtST + al(l-dsT)yS + a2( ~ ~ U2 )s Here tST is the scaled settling time for the ~im~ tion, Yos is the scaled maximum fractional overshoot relative to the setpoint change, and ( ~ ~u )s is the scaled sum of control moves over the simulation period. dST, tST ~ Yos and ~u all of course vary with each closed-loop simulation, but for clarity one has not explicitly indicated the dependence on case i in this equation. The wei~hting factors used for most experiments are ao: 0.75, al: 0-2, a2: 0-05-~ The overall cost function J is simply the sum of the individual Ji's:
J=~Ji i .
CA 022~2428 1998-10-20 W O 98100763 PCT~US97/09473 S
~ The scaled settling time tST is based on the computed settling time tST as follows. First, the maximum and minimum expected values are computed for tsT
max max max tST = ATp Tp + ATd d min min min tST = ATp Tp +ATd Td and then tST is linearly scaled within these limits, and bounded outside:
1 . 0 if tST > tST
t = ~ 0 0 if tST < tsT
m~ n '-ST '--ST
t~ax min otherwlse ST tsT
A similar approach is followed for calculating Yos from yO5. A fractional overshoot of 0.0 (i.e., no overshoot) maps to a Yos value of 0.0, and a fractional overshoot of 20 percent or more (relative to the setpoint change) maps to 1.0, with the scaling in 15 between again being linear. For the squared sum of control moves, these bounds are 1.0 and 3Ø
The above has specifled in some detail how the optimization algorithm evaluates a parameter vector. This scheme has been followed for neural network tuner designs for different controllers. Parameter choices for two types of controllers -- PI/PID and 20 "higher order linear controllers" (HOLC) are shown below:
CA 022~2428 1998-10-20 W O 98t00763 PCTrUS97/09473 Design PID/PI Tuner TIOLC Tuner Parameter rmin ~ O
rmax 0.25 0.5 Amin 0.2 0.1 Tp max 2.0 2.0 ATp min 2.0 1.0 ATd max 8.0 6.0 ATd Thus, both the PID/PI tuner and the HOLC tuner are designed to allow robustness/performance tradeoffs to be done by simply adjusting one knob (dST).
An outline of the evolutionary computing algorithm used for the design of the neural network tuners (the parameter vector is notated as w rather than p) is provided:
1. Generate an initial set or population of vectors wl, . . ., wN and evaluate cost function values Ji=J(Wi) max argmax ComputeJmax= i Ji, Wmax i i-2. Select a parameter vector w' at random (uniform distribution) from population 3. Select the perturbation standard deviation ~w randomly from between specif1ed min max 15 lower and upper bounds, C~ w and ~ w CA 022~2428 1998-10-20 W 098/00763 PCT~US97/09473 4. Generate a perturbation vector ~,~, ~N(0, ~w ) 5. Generate a parameter vector w" ~ w' + ~w and evaluate J(w'~
NONLINEAR-~PPRO,YIMATOR-BASED AUTOMATIC TUNER
BACKGROUND OF THE INVENTION
The present invention pertains to tuners for controllers, and particularly to automatic tuners for controllers. More particularly, the invention pertains to nonlinear-S approximator-type of tuners. U.S. Patent No. 5,311,421, by Masahide Nomura et al., issued May 10, 1994, and entitled "Process Control Method and System for Performing Control of a Controlled System by Use of a Neural Network," provides background information for the present invention and is hereby incorporated by reference in this description. Reference may also be had to German Patent DE 4433332, which describes the use of a neural network as a tuner for a controller.
SUMMARY OF THE INVENTION
The invention is a technique for developing a tuner which is used for tuning or optimally guiding a controller. The tuner has a preprocessor for transforming a set of input signals into a set of normalized parameters. These parameters are inputted to a nonlinear approximator that has been setup off-line without advance supervised learning using a plurality of closed-loop simulations. An optimization algorithm minimi7es the value of a cost function which is derived from the closed-loop simulations and which indicates the performance of the simulations, to produce optimized parameters of the nonlinear approximator. The nonlinear approximator outputs a set of normalized tuning parameters, derived on the basis of the minimi7ed value of the cost function, to a postprocessor which scales this set of parameters into controller tuning parameters which go to the controller. The nonlinear approximator can be based on neural networks or other parametrized nonlinear structures. In essence, deterrnining nonlinear approximator parameters actually amounts to designing or setting-up a tuner. In turn, the tuner tunes the controller having a closed loop, with appropriate controllerparameters.
BRIEF DESCRIPTION OF THE DRAWING
Figure 1 shows a basic controller and process relationship.
Figure 2 shows a controller with a tuner, along with a process.
Figure 3 shows a process with a step-function input.
Figure 4 shows the characteristics of an open loop step response of the process in figure 3.
~ENDED
., . .. -- ................ . . ....... . ... ..
CA 022~2428 1998-10-20 22/C~
Figure 5 shows a graph with a step tunction having a resultant overshoot and an oscillatory settling which has to be modeled as at least a second order system with delay.
Figure 6 reveals the closed-loop step response parameters of a system.
,~E~E~
. _ ~ . .
CA 022~2428 1998-10-20 W O 98/00763 PCTrUS97/09473 Figure 7 reveals a tuner with separate inputs for process characteristic parameters and closed-loop performance parameters.
Figure 8 is a graph of fast settling performance with overshoot.
Figure 9 is a graph of slow settling performance without overshoot.
S Figures l Oa and I Ob show an automatic tuner for processes that can be modeled as first-order linear systems with delay.
Figure 11 shows a neural network tuner used in conjunction with a parametrized neurocontroller.
Figure 12 illustrates a prior art neural network tuner with a supervised learning l O algorithm feature.
Figure 13 is a schematic of one framework of the present invention, not relying on supervised learning.
Figure 1 4a shows an example of a tuner for a proportional-integral (PI) controller embodied as a neural network with one tuning knob input and outputs for proportional and integral gains.
Figure 1 4b shows a tuner embodied as a computationally simple compositional mapping with one tuning.
Figure 15 shows how the tuner of figure 14 can be used for a known linear first-order process.
Figure 16 is a graph for slow or fast settling, dependent upon tuner adjustment input.
Figure 17 is a sçh~n ~tic for developing the tuner of figure 15 using a nonlinear optimization algorithm.
DESCRIPTION OF THE EMBODIMENT
Figure 1 shows a simple system proportional-integral-derivative (PID) controller12 that may be used for control of the t~ dlllre in a room. The setpoint Yd for temperature is input at line 13 to controller l 4 which has a u output on line 15 for controlling an airflow, such as heat or hot air flow, to process 16. The controller output goes to the furnace via a gas valve or other furnace controlling device. Process 16 is the room having its space le~ ld~ controlled. An output y on line 18 indicates the ten~ re of the room. Signal y is fed back to an input 20 of controller 14. In other words, controller 14 looks at the current temperature on line 18 and the setpoint CA 022~2428 1998-10-20 W O 98/00763 PCTrUS97/09473 (thermostat) temperature (y desired) on line 13, and uses the difference between the temperatures to calculate heating input 15 to the room or process 16.
"u" can be shown relative to various domains in a simple proportional-integral (PI) controller according to the following equations, where:
kc is the proportional gain for the controller;
k; is the integral gain for the controller;
t is time;
~ is an integration variable;
e(t) is the process output error at time t;
u(t) is the controller output at time t;
y(t) is the process output at time t;
yd(t) is the desired process output or setpoint of time t;
U(s) is the Laplace transform of u(t);
E(s) is the Laplace transform of e(t);
j is a summ:-tion index; and I is a discrete time index.
u(t) = kce(t) + k j Je(r)dr ( I ) where e(t) = Yd (t) - y(t) U(s) = kcE(s) + ' E(s) (2) u(l) = kce(l) + k, ~e(j)~t (3) j=o Equation (1) is stated for the continuous-time domain, equation (2) for the Laplace or frequency domain, and equation (3) for the discrete-time domain. The proportional (kc) and integral (kj) gains are set to values a~plopl;ate for the application 25 of control system 14 to process 16. ~t is the sampling interval.
CA 022~2428 1998-10-20 W 098/00763 PCTrUS97/09473 --4--The algorithms, used for setting controller gains, for instance kc and ki, are called "tuning algorithrns." The tuning algorithm is implemented in a tuner 21 of figure 2. The tuner puts gains kc and kj into controller 14 via line 22. To set these gains is often done by a "trial and error" approach in the related art. Current tuners in the art are 5 not wholly satisfactory.
Inputs to tuner 21 need to indicate two types of inforrnation. The first type ofinformation is the relevant dynamics of the process (e.g., characteristics of the room, such as its heat loss~ heating-up time constant, and the delay due to ductwork from the furnace to the room). The second type of information is desired dynarnics of the closed 10 loop (e.g., time to reach the setpoint, settling time, overshoot, tolerances of overshoot, and importance of reducing control energy).
Figure 3 shows a step function 24 of u to input 15 of process 16 and output y from line 18. Figure 4 shows the characteristics of an open loop step response of process 16. The process is of a first order with a delay linear system which is usually 15 sufficient for most industrial and building control processes. Such system has three process model parameters Kp, Tp and TD.
Thegainis K = (y,- yo)/(u,- 1~o) The time constant, T = t2 ~ t" (which is the time that y takes to reach 63 percent of the final output). The dead time or delay is T" = t,--to . If Kp, Tp, TD~ are changed, 20 then the curve for y will change.
Figure 5 shows a graph with a step function having a resultant overshoot and an oscillatory settling 25 which has to be modeled as at least a second order system with delay, relative to input 24.
Figure 1 shows a closed-loop system 12 with a step function signal 24 to input 13. Figure 6 reveals the closed-loop step response parameters of system 12.
Process overshoot is y~5 =
Settling time is t,.7 = t, - to Rise time is tpy~ = t2--to Control energy is Ec - ~(u(r) - u(r- 1))2 CA 022~2428 1998-10-20 W O 98/00763 PCTfUS97/09473 One important objective in tuning a controller is to manipulate these and other closed-loop response parameters. For example, large overshoots are typically undesirable~ but often unavoidable if fast settling times are required (the settling time is the time after the setpoint change after which the process output y is within some S quantity, e.g. 95%, of the desired output or setpoint Yd, with the percentage computed relative to the magnitude of the setpoint change). For fast settling, control energies will also typically be high. The response parameters have a complex relationship to each other and to the tuning parameters, and no formulae are available that can accurately characterize this relationship.
Different tuning methods take different types of tuner inputs of the relevant dynamics of the process and desired dynamics of the loop, and provide gains for different types of controllers. The present invention of tuning can be used for any kind of linear or nonlinear controller, not just a PID type. Tuner inputs can be customized as desired, and the tuner can be made robust to whatever extent desired. Robustness is the 15 graceful tolerance of uncertainty of knowledge about the system. It is also the graceful tolerance of change or drift in the system.
Two kinds of parameters are "nominal" and "estimated". For example, Kp is the nominal or actual gain and Kp is estimated gain. Controllers are designed to take into consideration that Kp ~ Kp.
Features of the present invention are the use of one or several nonlinear approximators in the tuner, and the use of a general optimization system. The following types of nonlinear approximators are explicitly considered:
~ Conventional multilayer perceptron neural networks ~ Computationally simple, compositional sigmoidal nonlinear mappings ~ Radial basis function networks ~ Functional link networks ~ Cerebellar model articulation controller (CMAC) networks ~ Fuzzy logic models ~ Wavelet networks ~ Polynomial expressions .. .. . . . .. . . .
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 ~ Specific nonlinear parametrized structures based on human intuition about controller tuning The common aspect of these and other relevant approximator structures is that they implement nonlinear mappings the form of which can be adjusted by manipulating 5 parameters associated with the approximator. Thus generally a nonlinear approximator, as used in this invention, can be expressed as follows:
Pc = F(I; Pa) where Pc is the vector of controller tuning parameters, I is the input vector to the nonlinear approximator, and Pa is the parameter vector for the approximator. For10 example, a conventional multilayer perceptron neural network may implement the mapping F as:
- PC~ +exp -~,w,lhj+b~ ) 15 where the hi are intermediate ("hidden unit") variables computed as:
[ k = 1 D
Here the parameter vector pa consists of the input-to-hidden weights w , the hidden-20 to-output weights wij, and the hidden and output bias weights bi and b3 . i and k are summation indices, j is the element of the approximator output vector, h is the number of hidden units, and n is the number of elements of the input vector I. These expressions relate to "one-hidden-layer" neural networks. Multiple layers of hidden units can also be used.
As another example, a computationally simple, compositional sigmoidal nonlinear mapping may implement F as:
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 pC~j] ~vii l+jI.wi + ail where the parameter vector Pa comprises the vectors wi and the scalars ai and vij. i is a summation index and b is the number of nonlinear elements in the structure. The period (.) denotes the dot product operator. Note that this form of the nonlinear approximator, which otherwise resembles a neural network, does not require the use of transcendental functions such as exponentials or hyperbolic tangents which are costly to compute. This approximator is the approach of first choice for this invention. Again, multiple "layers"
of the nonlinearities can be used.
As a third exarnple, a radial basis function network may implement the mapping Fas:
i=N 2 pClj] = ~ wjj exp~lI - ,Ui ll/~ ) j=, i where the parameter vector Pa comprises the vectors ,ui and the scalars ~i and wij. i is a summation index and N is the number of Gaussian nonlinearities in the structure.As a final example, a polynomial expansion may be used as the nonlinear approximator such as:
k=n k=r~=k pC~j]= ~aklk + ~ ~bklIkIl k=1 k=ll=l where the parameter vector Pa comprises the scalars ak and bkl and n is the number of elements in the input vector I. k and 1 are s~lmm~tion indices.
Figure 7 reveals a controller system 30 which has a nonlinear-approximator-based automatic tuner 21 having separate inputs for process characteristic parameters and closed loop performance parameters. The latter can be used to adjust closed-loop performance for slow or fast settling. Inputs 26 and 27 permit one pre-designed tuner 21 to be used for many applications or be universal for large and various classes of applications. A nonlinear approximator 28 is optimized off-line but without using supervised learning methods. Nonlinear approximator 28 may be a neural network but is not used as a nonlinear regression model (as it is in Nomura et al. -- see their Fig. 33).
.. ..... . . . . . . ... . . .. .. .
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 In addition to neural network or other nonlinear approximator 28, automatic tuner 21 can contain preprocessing and post-processing functions. These include scaling functions which are done outside of the neural network. Scaling is performed mainly for linearity purposes. Process characteristic parameters 26 (notated as pp later) 5 and closed-loop performance parameters 27 (notated as pj later) allow one pre-designed automatic tuner 21 to be used for a broad variety of applications. No on-line orapplication-specific training or optimization of the neural network, or modification of other components of the automatic tuner, are needed.
A specific example is neural-network-based automatic tuner 21 for processes 10 that can be modeled as first-order with dead time linear processes. Such processes can be modeled with three process model (or process characteristic) parameters 26: the process gain, time constant, and dead time. One closed-loop performance parameter is also assumed: a settling time knob dST (this constitutes the closed loop performance parameter at 27 in figure 7), which can be used to adjust closed-loop performance for 15 slow or fast settling. There may be other closed-loop performance parameters on input 27. In this case, the neural network can be off-line optimized without using supervised learning that is used in the related art (as discussed below) using a simulation-based optimization system.
Figure 8 shows the fast settling (i.e., low tsT) performance with overshoot for a low dsT input 27. Figure 9 shows the slow settling (i.e., high tST) performance without overshoot for a high dST input 27. There may be a knob for dST that provides an analog-like variation, such as from zero to one, to automatic tuner 21.
The key feature of the present invention is the design of nonlinear approximator28. During operation in a setup similar to figure 7, automatic tuner 21 can have a basic architecture, with pre-processor, non-linear approximator 28 and post-processor, as shown in figure I Oa. Note in figure 1 Ob that three process characteristic parameters are input to automatic tuner 21: estimates of process gain, time constant, and dead time ,~ A ,~
(notated K~, T~, and T~, ). Output 22 of the tuner 21 consists of controller parameters (for example, for a PID controller, these would be the proportional, integral, and 30 derivative gains) PC In figure 10b, phias iS an internal variable to automatic tuner 21 that consists of baseline controller parameters which are then modified in a context-. . .
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 sensitive way by neural network or a nonlinear approximator 28. Note that by intelligently structuring autotuner 21, the number of inputs required for nonlinear approximator 28 is reduced -- in this case, to two. T" / T to line 46 and dST to line 29 are inputted to nonlinear approximator 28. An output 31, PC, which is a control S parameter, is fed into s~-mming junction 32. A pc bias is added at line 33 to Pc at summer 32 which outputs a pcnom, which is the vector of nominal (unscaled) control parameters, at line 34 and is inputted to scaling mechanism 35. Output 22 of scaling 35 goes to controller 14.
Tuner 21 can be used in control structure having both linear and nonlinear l 0 components. For example, tuner 21 can be used in conjunction with a parametrized neurocontroller 36, as shown in figure 11.
Neural network based automatic tuner 21 is optimized off line using a design framework. An "evolutionary computing" algorithm has been developed that incorporates aspects of genetic algorithms. Also, a gradient search can be enabled, with 15 gradients numerically computed.
The flexible nature of the design framework perrnits one to optimize for criteria that more conventional approaches seldom permit. In particular, performance criteria need not be quadratic, or even differentiable; control structures can be arbitrarily nonlinear; and robustness can explicitly be considered. The present approach to the 20 optimization of neural network based automatic tuner 21 does not require "learning in advance" in which input-output combinations must be first compiled using a separator optimization program, and neural network 28 then trained using supervised learning.
The present approach can thus be contrasted to the approach of Nomura et al. (U.S.
Patent 5,311,421) in which the use of a neural network for tuning is also disclosed but in 25 which the development of the neural network requires a supervised learning algorithm for which a~lopllate teaching data must be collected.
One may use supervised learning of Nomura et al. for neural network tuner 37 of figure 12. For tuner 37, inputs may be ~lesign~ted x and outputs designated Pc~ dST is part of x. Pc are control parameters. Pc is compared with the desired output Pc* and the 30 result of the comparison at comparator junction 38 is error ~, which is fed to learning algorithm 39. Algorithm 39 then provides an output to neural network 40 to adjust the .. .. . ..... . .... .. ...
CA 022~2428 1998-10-20 2/~ o/~
lo-weights for output adjustments. For that supervised learning, a large database of (Y, Pc*) pairs need to be compiled in advance. Pc* needs to be computed for many different cases in advance. Nomura et al. involves the inputting of inforrnation with the characteristics contained therein to a neural network 40 for "learning in advance a correlation" between the information containing the characteristics and a control parameter and determining a control parameter for a controller. "Learning in advance"
means the deterrninin~ of optimal control parameters corresponding to different characteristics of the model of the combined controller-controlled system and then using the optimal control parameters as learning teacher data. In Nomura et al., it is also assumed that input-output combinations are available as learning data (col. 11, lines 1 1-17). The Cj's in equation (10) are ideal outputs that the neural network is trained to match, and thus Cj's are to be available for neural network training. The neural network is used as a nonlinear regression model and thus the neural network is regressed using input-output training data.
l S The present approach does not rely on such supervised le~rning No collection of (x, Pc*) pairs is needed and no computation of (optimal) Pc*~s is required. The present neural network is developed using off-line optimi7~tion without any need for "learning in advance" in the sense of Nomura et al. No input-output training data is needed.
It should be noted that the prior art of Nomura et al., by virtue of its reliance on a supervised learning approach, is restricted to certain types of neural network tuners.
Our invention eschews supervised learning entirely in favor of a simulation-based optimization approach, and an important ancillary benefit of this approach is that the tuner can comprise ;~bil~ nonlinear approximators, including but certainly not limited to the type of neural network structure employed in the cited prior art.German Patent DE 4433332 also describes the use of a neural network as a tuner for a PID controller. As in Nomura et al., learning data for the neural network must first be produced before the neural network itself can be optimized.
Figure 13 shows the design framework 41 of the present invention schematically. For each parameter vector p, closed-loop simulations are run and a performance criterion value computed. No learning teacher data is required. The vector p can contain gains for a linear control structure, so that one can optimize a fixed controller for some criterion; it can contain parameters for a nonlinear-approximator-o~~s~
CA 02252428 1998-10-20 ~2/C6/~/~
-IOA
based tuner or neural network controller (e.g., pararnetrized neurocontroller (PNC) 36 in figure 11); or it can contain pararneters for both the tuner ~nd controller, so that both O~) S~
.. . .. . . . ..
CA 022~2428 1998-10-20 modules can simultaneously be optimized. The objective of the design activity is to develop tuners or controllers that can be used for a broad range of applications. Thus, numerous closed-loop simulations must be run for every choice of p. In these simulations, various parameters are varied depending on the design requirements:5 ~ Process model parameters pp which are input to the process model (PM) in the closed-loop simulation ~ Estimation errors between these process model parameters and the process model parameter estimates pr, of which the latter are input to the tuner and/or controller l O ~ Initial conditions, especially in the case that the closed-loop controller or process model is nonlinear ~ Setpoint profiles, also for nonlinear controllers or nonlinear process models ~ Parameters PJ that affect the cost function--the use of these parameters allow us to develop one control solution for a class of parametrized criteria Each closed-loop simulation results in one elementary cost function Jj being computed. The results of all the simulations conducted for one vector p are composed into an overall cost function J which is returned to the optimization algorithm:
20 J = fJ(Jl~ J2, ~ ~, JN) where N is the number of closed-loop simulations performed (typically l O00).
The above paragraphs describe the general framework of the design system. For this project, one has made specific choices for optimization runs. Note these below:
~ Process models are first order with dead time. The process gain Kp is norm~li7~1 to 1.0, the process time constant Tp to 10Ø The process dead time Td varies so that the constraint in the next statement holds:
CA 022~2428 1998-10-20 ~ Estimates of Kp, Tp, and Td are input to the neural network controller or tuner.
These estimates are perturbations of the true model parameters as discussed below. Td is limited to the range [0,2 T ].
5 ~ The parameter estimate for a parameter Pi is bounded uniformly within [( I -r)pi, (l+r)pi], where r is a parameter. r is a function ofthe criterion parameter dST
which can vary between 0 and 1. This function is a linear function so that the minimum value of r, achieved when dST is 0, is rmin~ and the maximum value, corresponding to dST equal to 1, is rmaX~ Thus, dST controls the robustness of the control, in addition to being a settling time knob as discussed below.
~ The only criterion parameter employed is dST. The elementary criterion is:
Ji - aOdSTtST + al(l-dsT)yS + a2( ~ ~ U2 )s Here tST is the scaled settling time for the ~im~ tion, Yos is the scaled maximum fractional overshoot relative to the setpoint change, and ( ~ ~u )s is the scaled sum of control moves over the simulation period. dST, tST ~ Yos and ~u all of course vary with each closed-loop simulation, but for clarity one has not explicitly indicated the dependence on case i in this equation. The wei~hting factors used for most experiments are ao: 0.75, al: 0-2, a2: 0-05-~ The overall cost function J is simply the sum of the individual Ji's:
J=~Ji i .
CA 022~2428 1998-10-20 W O 98100763 PCT~US97/09473 S
~ The scaled settling time tST is based on the computed settling time tST as follows. First, the maximum and minimum expected values are computed for tsT
max max max tST = ATp Tp + ATd d min min min tST = ATp Tp +ATd Td and then tST is linearly scaled within these limits, and bounded outside:
1 . 0 if tST > tST
t = ~ 0 0 if tST < tsT
m~ n '-ST '--ST
t~ax min otherwlse ST tsT
A similar approach is followed for calculating Yos from yO5. A fractional overshoot of 0.0 (i.e., no overshoot) maps to a Yos value of 0.0, and a fractional overshoot of 20 percent or more (relative to the setpoint change) maps to 1.0, with the scaling in 15 between again being linear. For the squared sum of control moves, these bounds are 1.0 and 3Ø
The above has specifled in some detail how the optimization algorithm evaluates a parameter vector. This scheme has been followed for neural network tuner designs for different controllers. Parameter choices for two types of controllers -- PI/PID and 20 "higher order linear controllers" (HOLC) are shown below:
CA 022~2428 1998-10-20 W O 98t00763 PCTrUS97/09473 Design PID/PI Tuner TIOLC Tuner Parameter rmin ~ O
rmax 0.25 0.5 Amin 0.2 0.1 Tp max 2.0 2.0 ATp min 2.0 1.0 ATd max 8.0 6.0 ATd Thus, both the PID/PI tuner and the HOLC tuner are designed to allow robustness/performance tradeoffs to be done by simply adjusting one knob (dST).
An outline of the evolutionary computing algorithm used for the design of the neural network tuners (the parameter vector is notated as w rather than p) is provided:
1. Generate an initial set or population of vectors wl, . . ., wN and evaluate cost function values Ji=J(Wi) max argmax ComputeJmax= i Ji, Wmax i i-2. Select a parameter vector w' at random (uniform distribution) from population 3. Select the perturbation standard deviation ~w randomly from between specif1ed min max 15 lower and upper bounds, C~ w and ~ w CA 022~2428 1998-10-20 W 098/00763 PCT~US97/09473 4. Generate a perturbation vector ~,~, ~N(0, ~w ) 5. Generate a parameter vector w" ~ w' + ~w and evaluate J(w'~
6. If J(~ < Jmax then Replace wmay with w' Set w' to w"
Recompute Jmax and Wmax Goto Step 5 {Reapply perturbation}
Recompute Jmax and Wmax Goto Step 5 {Reapply perturbation}
7. Goto Step 2 According to the present invention, one can design a tuner by employing a nonlinear optimization algorithm in conjunction with a closed-loop simulation system.
The nonlinear optimization algorithm attempts to minimi7~ the value of the cost function J(p) by ~proJ~liately adjusting the nonlinear approximator parameters p.
Figures 1 4a and 1 4b show two similar nonlinear-approximator-based tuners 21. In Figure 1 4a, the nonlinear approximator is implemented as a multilayer perceptron neural network. The parameters to be optimized in this case are w1, . . ., wl4. In 20 Figure 1 4b, the nonlinear approximator is implemented as a computationally simple compositional mapping--the figure shows the mathematical formulation of the mapping.
Because of its computational simplicity, this is a preferred embodiment. The parameters to be optimized in this case are wl, w2, . . ., wb, al, a2, , ab, v11, V21, ~, vbl, V12, V22, ~ ~, Vb2- It should be noted that the implementations of Figures 14a and 14b have very similar features, and in both cases output KC and Kl as functions of dST and the respective parameters. However, the supervised learning algorithm described in Nomura et al. cannot directly be used for the approximator of Figure 1 4b.
Once the applopl;ate parameters are determined so as to minimi7~. J(p), tuner 21 can be applied to real systems (provided that the constraints assumed in the design hold for the 30 real system).
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 The optimization involved in tuner design may seem complex and time consuming. However, once designed, tuner 21 can be easily used in a variety of different applications on simple computational platforms. Figure 15 shows how tuner 21 is used for system control. Tuners can be designed and used in different 5 embodiments, depending on the nature of user-control of closed-loop performance desired, and the types of systems and controllers the tuner will be used with.
In one example embodiment of a fixed first order linear process, PI controller, for settling time tuning, one can describe how this invention can be applied to design a nonlinear-approximator-based tuner 21 for PI controller 14 with a "settling time" tuning 10 knob 29 when the processes for which tuner 21 is designed are assumed to be adequately characterized by first-order linear dynamics which are accurately known. A
system is illustrated in figure 15. The process model is expressed in the Laplace domain as:
15 Y(s) = ~ U(s) =~ y(t)+ " u(t) Ts+ 1 dt T T
P P P
where Kp is the process gain and Tp the time constant, the values for which are known (through experiments or process knowledge). Tuner 21 has one tuning knob 29, dSt, 20 which can be varied over a range 0.0 to 1.0 to make the process response to a setpoint change be fast or slow. For fast settling, one is relatively tolerant of overshoot in the response, and control action. For slow settling, however, one wants overshoot to be minimi7ed and control action not to be aggressive. Figure 16 shows curve 42 of y when dSt is low and curve 43 when dSt is high for setpoint change 24. A possible cost25 function 44 (figure 17) for this embodiment is the following:
J(w)= ((I-d5t)fst(tst)+ d5,foS(yos)+dstfu,~(~umay)) (1) In this expression, E(-) denotes the expectation operator, and dSt~[0,1~ denotes that the 30 expectation is to be taken over a (uniformly distributed) range from 0 to I of values of CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 dSt. tst represents the settling time of the closed-loop simulation, Yos the overshoot in the response, and ~umax the maximum control move encountered. l he functions fst(.), fos(-), and fum(-) are scaling functions so that the effect of the three terms is in accordance with the relative importance of each of these features. Note that since dSt is S an input to tuner 21 which outputs PI gains that influence the closed-loop behavior (and hence Yos, tst, and ~umax), the value of the expression (1 ) is a function of parameters p of the nonlinear approximator.
Once a tuner 21 is designed to minimi7e J(w) in equation (1), it can directly be used to control the target process. The control10 perforrnance can be adjusted through the use of the tuning knob dSt 29. When dSt is low (near 0), setpoint changes will be rapidly tracked. When dSt is high (near 1), fast setpoint tracking is no longer important, but output overshoot and the maximum actuator move will be minimi7ed Of course, since tuner design 21 is based on a specific process model, the tuner performance will largely be dependent on the accuracy 15 of the model. For cases where the model is not precisely known, tuners can be optimized for robust performance.
Equation (1 ) cannot be analytically solved in most cases of practical interest.However, it can be approxim~ted by Monte Carlo simulations to as high a degree of accuracy as required. This simulation approximation is guided by a nonlinear 20 optimization algorithm 45 of figure 17. Of course, high precision is gained at the expense of computing time.
For the embodiment of an uncertain first order with delay linear process, PID
controller, settling time tuning, one no longer assumes that the process is precisely known. Instead, tuner 21 is designed to o~ performance over a range of process 25 models. One extends the process model ofthe embodiment of a fixed first order linear process and PI controller, by adding a delay or dead-time:
Kpe-Td Tps+l 30 where Td is the process dead time. The three process parameters are assumed to lie within known spaces K (for Kp), T (for Tp), and ~) (for Td). In addition, one is now CA 022~2428 l998-l0-20 WO 98/00763 PCT~US97/09473 interested in a tuner for a PID controller in interactive form. The controller transfer function is:
() K ~,1+ ~ l+~"s ) E() T,s l + ~x~,/s s where U(s) is the controller output (Laplace transformed), Ti is the integral time, Td the derivative time, (x the rate amplitude (a known constant), and E(s) the error (Laplace transformed).
In this case, one can adopt the following cost function:
J(W) = d [0 1~ d~ )f" (tS~)+ d,"fo~(yO~)+ d~f~ Uma1~)) (2) K,~ ~K
Tp ~T
T~
Here, instead of minimi7ing an expectation, one is interested in minimi7ing the worst case performance. The maximum is computed over spaces for d5l, Kp, Tp, and T~. The 15 robustness associated with a tuner that is optimi~ed for Cost Function (2) implies a loss of nominal performance. If in fact the process model parameter estimates were known more precisely than one assumes here, then better control could be achieved. This suggests that a tuning knob could be added that allows the user of the tuner to make the robustness versus nominal perforrnance tradeoff on a case-by-case basis without 20 reoptimization. The next embodiment description discusses this aspect of the invention.
In an embodiment of an uncertain first order linear process, multiple lead-lag controller, for settling time and robustness tuning, the controller is not of the PID
variety but consists of three lead lag terrns:
CA 022~2428 1998-10-20 W O 98/00763 rCT~US97/09473 U _ ao (a1s+1) (a2s+l) (a3s+1) E
(s) - (bl s+l ) (b2s+1 ) (b3s+1 ) ( ) A second tuning knob, r, is added that also varies between 0 and 1. The parameter spaces over which tuner 21 operates are now a function of r:
J(w) = [ I] ((1 - d 5t )f5, (t 5, ) + d 5, fcs (Y OS ) + d st film (~Un~ax )) (3) r ~[0,1]
K" ~k:(r) T" ~ T(r) For simplicity, one may revert to the simple delay-less process model.
For r=0, the expectation is evaluated over small spaces K and T in this case, and 10 one is interested in maximi7in~ performance at the expense of robustness. As r is increased, the ranges increase in accordance with prespecified functions K(r) and T(r).
Note that the effect of r can be incorporated into dSt itself--the slower settling times associated with high dSt values are consistent with increased robustness. Thus, Cost Function (3) could be simplified by removing r and making K and T dependent on dSt.
In the embodiments discussed so far, tuner 21 input has not included any information on process 16 itself. The tuners can be used for different processes to the extent that they fall within the robustness spaces K, T, and (~) above. The next two embodiments are described in which the tuner is also provided process parameter estimates as inputs, and is thereby generically applicable to a considerably larger space 20 of processes. For the first embodiment, the generic feature is obtained simply by scaling controller gains.
In a process-generic tuner, first order process model, one may assume the designconditions of the embodiment of a first order linear process with a PI controller, and let the tuner be designed for specific nominal values of Kp and Tp. Let these values be 25 Kp and T , respectively. For a given setting of dSt, let the gains output by the tuner be Kc and K i ~ The same tuner outputs can be used for a different process . .
CA 022~2428 1998-10-20 (one that permits a first order approximation) with a gain of Kp and a time constant of TneW by adapting the gains as follows:
K K
K~l~'lV C p p Tnom T new e~-~ i P
p Scaling formulae can be derived for all cases where both the controller and the process model are linear. Similar formulae can also be derived for some cases where 10 the controller is nonlinear and the process model is linear. Such scaling formulae can compensate for arbitrary Kp and Tp values. In cases where the process model contains additional parameters, the nominal values above need to be functions of model parameters.
One may extend the embodiment of the process-generic tuner, first order process 15 model, by adding a delay (dead time) to the process model. To design a tuner that will work for all processes that can be modeled as a first-order linear system with delay, provided that the delay does not exceed some constant factor ~ of the time constant Tp, one can provide a process model input to the tuner. This input is Tr = Td The design cost function (1) is now extended as:
J(w) = [ 1] ((1 - d 5, )f5, (tS, ) + d 5t fO5 (yO5 ) + d st fum (~umax )) (4) Tr ~~,~]
So the tuner design is perforrned for fixed values of Kp and Tp and with Td varying between 0 and ~Tp. Once the design is complete--a weight vector w is determined that CA 022~2428 1998-10-20 minimi7~s Cost Function (4)--the resulting tuner can then be used for any process that can be modeled as first order with dead time, the only constraint imposed being that the dead time is no greater than ~ times the process time constant. For such use, estimated values of process gain, time constant, and dead time need to be input to the tuner. These 5 are used to compute Tr and also to effect scalings such as those defined above. Since Cost Function (4) does not address robustness, these parameters must be accurately known.
In a state feedback controller, having a nonlinear process model and input disturbance rejection tuning, one assumes that the process is nonlinear, with model structure known:
x, (t + 1)' ~2xl (t)x2 (t) + x, 2 (t) ~ 0' x2 (t + 1) = xl (t) + x3(t) + 0 u(t) (5) ~x3(t + l)~ ~ sin(~c2(f)) J ~yJ
y(t) = x1 (t) In this model, the xj are state variables and ~ is a parameter. For a control structure, one adopts a linear state-feedback control law:
u(t) = a1~x1(t) + a2~x2(t) + a3~x3(t) where ~xj = x(i) - xj, the difference between the desired state value and the measured value. Thus, the tuning problem in this case is to estimate ~pprol)l;ate values of the controller parameters al, a2, and a3. This is to be done on the basis of two knobs: One is a performance criterion setting ddr that allows the user to tradeoff control action 25 versus fast disturbance rejection for step input disturbances. The other is an estimate y of the model parameter ~. By incorporating this latter parameter, one allows the same controller and tuner to be used over a range of processes of structure similar to equation (5) without being limited to process models with ~ equal to some specific nominal value. The fact that we use an estimate of ~ instead of the actual value implies that the .... . . ~ ~
CA 022~2428 1998-10-20 W O 98100763 PCTnUS97109473 tunings will have designed-in robustness to errors in parameter estimation. A suitable cost function for this embodiment is the following:
J(w) = dd, ~ [0,1]
[r (ddr),r + ~(dd )] (( dr)fdr(tdr) + ddrf (~(~u))) (6) d ~ [-d m~X ~ dma~ ]
X(O) ~ Xo Values of ~ for which this tuner is designed, range from 1 to 5. The estimation error for tuning input purposes is up to ~t ~, and is a function of ddr so that greater robustness is demanded for slow rejection time. The step input disturbances d that one is interested in rejecting have amplitudes between -dmaX and dmaX. tdr is the time to reject the input disturbance and ~(~u) is the sum-squared control action. The expectation must also be computed over a space Xo of initial conditions x(0).
For the embodiment of a state feedback controller having a nonlinear process model and input disturbance tuning as above, one can also employ a nonlinear controller of known structure, for example:
u(t) = a1Xl(t) + a2X2(t) + a3x3(t) + a4Xl(t)X2(t) + asxl(t) x3(t) + a6x2(t)x3(t) In this embodiment the tuner outputs six parameters a1 - a6. Cost Function (6) can again be used.
For a PD controller, integrating linear process, easy-to-compute response feature inputs, rise time tuning, the following may be implemented.
The process model is:
s(Tps+l) The controller is:
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 e(t) - e(t - ~t) u(t) = Kce(t) + Kd At Tuner outputs are proportional gain Kc and derivative gain Kd. The tuner input 5 consists of simple features computed from the process output after exciting the process at time to with a pulse of height H and width W. The process output is measured at times to+n 4 , for n = O, 1, . . ., 10. Ten features are computed simply as y(to+n 4 ) -y(to) for n = 1, 2,...7 10. Neither Tp nor Kp are assumed to be known in thisembodiment. However, appropriate values for H and W do depend on the process, and 10 some general notions of the time constant and gain are needed.
In all the above embodiments, the key piece of the tuner is a nonlinear approximator. Various nonlinear approximators can be used in this context, including multilayer perceptron neural networks, computationally simple sigmoidal compositions, radial basis function networks, functional link networks, CMAC networks, fuzzy logic 15 models that employ fu~ification and/or defuzzification and/or membership functions, wavelet networks, polynomial expansions, and specific nonlinear parametrized structures.
The nonlinear optimization algorithm attempts to minimi7~ the value of the cost function J(p) by ~proJ~liately adjusting the nonlinear approximator parameters p.
Figures 1 4a and 1 4b show two similar nonlinear-approximator-based tuners 21. In Figure 1 4a, the nonlinear approximator is implemented as a multilayer perceptron neural network. The parameters to be optimized in this case are w1, . . ., wl4. In 20 Figure 1 4b, the nonlinear approximator is implemented as a computationally simple compositional mapping--the figure shows the mathematical formulation of the mapping.
Because of its computational simplicity, this is a preferred embodiment. The parameters to be optimized in this case are wl, w2, . . ., wb, al, a2, , ab, v11, V21, ~, vbl, V12, V22, ~ ~, Vb2- It should be noted that the implementations of Figures 14a and 14b have very similar features, and in both cases output KC and Kl as functions of dST and the respective parameters. However, the supervised learning algorithm described in Nomura et al. cannot directly be used for the approximator of Figure 1 4b.
Once the applopl;ate parameters are determined so as to minimi7~. J(p), tuner 21 can be applied to real systems (provided that the constraints assumed in the design hold for the 30 real system).
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 The optimization involved in tuner design may seem complex and time consuming. However, once designed, tuner 21 can be easily used in a variety of different applications on simple computational platforms. Figure 15 shows how tuner 21 is used for system control. Tuners can be designed and used in different 5 embodiments, depending on the nature of user-control of closed-loop performance desired, and the types of systems and controllers the tuner will be used with.
In one example embodiment of a fixed first order linear process, PI controller, for settling time tuning, one can describe how this invention can be applied to design a nonlinear-approximator-based tuner 21 for PI controller 14 with a "settling time" tuning 10 knob 29 when the processes for which tuner 21 is designed are assumed to be adequately characterized by first-order linear dynamics which are accurately known. A
system is illustrated in figure 15. The process model is expressed in the Laplace domain as:
15 Y(s) = ~ U(s) =~ y(t)+ " u(t) Ts+ 1 dt T T
P P P
where Kp is the process gain and Tp the time constant, the values for which are known (through experiments or process knowledge). Tuner 21 has one tuning knob 29, dSt, 20 which can be varied over a range 0.0 to 1.0 to make the process response to a setpoint change be fast or slow. For fast settling, one is relatively tolerant of overshoot in the response, and control action. For slow settling, however, one wants overshoot to be minimi7ed and control action not to be aggressive. Figure 16 shows curve 42 of y when dSt is low and curve 43 when dSt is high for setpoint change 24. A possible cost25 function 44 (figure 17) for this embodiment is the following:
J(w)= ((I-d5t)fst(tst)+ d5,foS(yos)+dstfu,~(~umay)) (1) In this expression, E(-) denotes the expectation operator, and dSt~[0,1~ denotes that the 30 expectation is to be taken over a (uniformly distributed) range from 0 to I of values of CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 dSt. tst represents the settling time of the closed-loop simulation, Yos the overshoot in the response, and ~umax the maximum control move encountered. l he functions fst(.), fos(-), and fum(-) are scaling functions so that the effect of the three terms is in accordance with the relative importance of each of these features. Note that since dSt is S an input to tuner 21 which outputs PI gains that influence the closed-loop behavior (and hence Yos, tst, and ~umax), the value of the expression (1 ) is a function of parameters p of the nonlinear approximator.
Once a tuner 21 is designed to minimi7e J(w) in equation (1), it can directly be used to control the target process. The control10 perforrnance can be adjusted through the use of the tuning knob dSt 29. When dSt is low (near 0), setpoint changes will be rapidly tracked. When dSt is high (near 1), fast setpoint tracking is no longer important, but output overshoot and the maximum actuator move will be minimi7ed Of course, since tuner design 21 is based on a specific process model, the tuner performance will largely be dependent on the accuracy 15 of the model. For cases where the model is not precisely known, tuners can be optimized for robust performance.
Equation (1 ) cannot be analytically solved in most cases of practical interest.However, it can be approxim~ted by Monte Carlo simulations to as high a degree of accuracy as required. This simulation approximation is guided by a nonlinear 20 optimization algorithm 45 of figure 17. Of course, high precision is gained at the expense of computing time.
For the embodiment of an uncertain first order with delay linear process, PID
controller, settling time tuning, one no longer assumes that the process is precisely known. Instead, tuner 21 is designed to o~ performance over a range of process 25 models. One extends the process model ofthe embodiment of a fixed first order linear process and PI controller, by adding a delay or dead-time:
Kpe-Td Tps+l 30 where Td is the process dead time. The three process parameters are assumed to lie within known spaces K (for Kp), T (for Tp), and ~) (for Td). In addition, one is now CA 022~2428 l998-l0-20 WO 98/00763 PCT~US97/09473 interested in a tuner for a PID controller in interactive form. The controller transfer function is:
() K ~,1+ ~ l+~"s ) E() T,s l + ~x~,/s s where U(s) is the controller output (Laplace transformed), Ti is the integral time, Td the derivative time, (x the rate amplitude (a known constant), and E(s) the error (Laplace transformed).
In this case, one can adopt the following cost function:
J(W) = d [0 1~ d~ )f" (tS~)+ d,"fo~(yO~)+ d~f~ Uma1~)) (2) K,~ ~K
Tp ~T
T~
Here, instead of minimi7ing an expectation, one is interested in minimi7ing the worst case performance. The maximum is computed over spaces for d5l, Kp, Tp, and T~. The 15 robustness associated with a tuner that is optimi~ed for Cost Function (2) implies a loss of nominal performance. If in fact the process model parameter estimates were known more precisely than one assumes here, then better control could be achieved. This suggests that a tuning knob could be added that allows the user of the tuner to make the robustness versus nominal perforrnance tradeoff on a case-by-case basis without 20 reoptimization. The next embodiment description discusses this aspect of the invention.
In an embodiment of an uncertain first order linear process, multiple lead-lag controller, for settling time and robustness tuning, the controller is not of the PID
variety but consists of three lead lag terrns:
CA 022~2428 1998-10-20 W O 98/00763 rCT~US97/09473 U _ ao (a1s+1) (a2s+l) (a3s+1) E
(s) - (bl s+l ) (b2s+1 ) (b3s+1 ) ( ) A second tuning knob, r, is added that also varies between 0 and 1. The parameter spaces over which tuner 21 operates are now a function of r:
J(w) = [ I] ((1 - d 5t )f5, (t 5, ) + d 5, fcs (Y OS ) + d st film (~Un~ax )) (3) r ~[0,1]
K" ~k:(r) T" ~ T(r) For simplicity, one may revert to the simple delay-less process model.
For r=0, the expectation is evaluated over small spaces K and T in this case, and 10 one is interested in maximi7in~ performance at the expense of robustness. As r is increased, the ranges increase in accordance with prespecified functions K(r) and T(r).
Note that the effect of r can be incorporated into dSt itself--the slower settling times associated with high dSt values are consistent with increased robustness. Thus, Cost Function (3) could be simplified by removing r and making K and T dependent on dSt.
In the embodiments discussed so far, tuner 21 input has not included any information on process 16 itself. The tuners can be used for different processes to the extent that they fall within the robustness spaces K, T, and (~) above. The next two embodiments are described in which the tuner is also provided process parameter estimates as inputs, and is thereby generically applicable to a considerably larger space 20 of processes. For the first embodiment, the generic feature is obtained simply by scaling controller gains.
In a process-generic tuner, first order process model, one may assume the designconditions of the embodiment of a first order linear process with a PI controller, and let the tuner be designed for specific nominal values of Kp and Tp. Let these values be 25 Kp and T , respectively. For a given setting of dSt, let the gains output by the tuner be Kc and K i ~ The same tuner outputs can be used for a different process . .
CA 022~2428 1998-10-20 (one that permits a first order approximation) with a gain of Kp and a time constant of TneW by adapting the gains as follows:
K K
K~l~'lV C p p Tnom T new e~-~ i P
p Scaling formulae can be derived for all cases where both the controller and the process model are linear. Similar formulae can also be derived for some cases where 10 the controller is nonlinear and the process model is linear. Such scaling formulae can compensate for arbitrary Kp and Tp values. In cases where the process model contains additional parameters, the nominal values above need to be functions of model parameters.
One may extend the embodiment of the process-generic tuner, first order process 15 model, by adding a delay (dead time) to the process model. To design a tuner that will work for all processes that can be modeled as a first-order linear system with delay, provided that the delay does not exceed some constant factor ~ of the time constant Tp, one can provide a process model input to the tuner. This input is Tr = Td The design cost function (1) is now extended as:
J(w) = [ 1] ((1 - d 5, )f5, (tS, ) + d 5t fO5 (yO5 ) + d st fum (~umax )) (4) Tr ~~,~]
So the tuner design is perforrned for fixed values of Kp and Tp and with Td varying between 0 and ~Tp. Once the design is complete--a weight vector w is determined that CA 022~2428 1998-10-20 minimi7~s Cost Function (4)--the resulting tuner can then be used for any process that can be modeled as first order with dead time, the only constraint imposed being that the dead time is no greater than ~ times the process time constant. For such use, estimated values of process gain, time constant, and dead time need to be input to the tuner. These 5 are used to compute Tr and also to effect scalings such as those defined above. Since Cost Function (4) does not address robustness, these parameters must be accurately known.
In a state feedback controller, having a nonlinear process model and input disturbance rejection tuning, one assumes that the process is nonlinear, with model structure known:
x, (t + 1)' ~2xl (t)x2 (t) + x, 2 (t) ~ 0' x2 (t + 1) = xl (t) + x3(t) + 0 u(t) (5) ~x3(t + l)~ ~ sin(~c2(f)) J ~yJ
y(t) = x1 (t) In this model, the xj are state variables and ~ is a parameter. For a control structure, one adopts a linear state-feedback control law:
u(t) = a1~x1(t) + a2~x2(t) + a3~x3(t) where ~xj = x(i) - xj, the difference between the desired state value and the measured value. Thus, the tuning problem in this case is to estimate ~pprol)l;ate values of the controller parameters al, a2, and a3. This is to be done on the basis of two knobs: One is a performance criterion setting ddr that allows the user to tradeoff control action 25 versus fast disturbance rejection for step input disturbances. The other is an estimate y of the model parameter ~. By incorporating this latter parameter, one allows the same controller and tuner to be used over a range of processes of structure similar to equation (5) without being limited to process models with ~ equal to some specific nominal value. The fact that we use an estimate of ~ instead of the actual value implies that the .... . . ~ ~
CA 022~2428 1998-10-20 W O 98100763 PCTnUS97109473 tunings will have designed-in robustness to errors in parameter estimation. A suitable cost function for this embodiment is the following:
J(w) = dd, ~ [0,1]
[r (ddr),r + ~(dd )] (( dr)fdr(tdr) + ddrf (~(~u))) (6) d ~ [-d m~X ~ dma~ ]
X(O) ~ Xo Values of ~ for which this tuner is designed, range from 1 to 5. The estimation error for tuning input purposes is up to ~t ~, and is a function of ddr so that greater robustness is demanded for slow rejection time. The step input disturbances d that one is interested in rejecting have amplitudes between -dmaX and dmaX. tdr is the time to reject the input disturbance and ~(~u) is the sum-squared control action. The expectation must also be computed over a space Xo of initial conditions x(0).
For the embodiment of a state feedback controller having a nonlinear process model and input disturbance tuning as above, one can also employ a nonlinear controller of known structure, for example:
u(t) = a1Xl(t) + a2X2(t) + a3x3(t) + a4Xl(t)X2(t) + asxl(t) x3(t) + a6x2(t)x3(t) In this embodiment the tuner outputs six parameters a1 - a6. Cost Function (6) can again be used.
For a PD controller, integrating linear process, easy-to-compute response feature inputs, rise time tuning, the following may be implemented.
The process model is:
s(Tps+l) The controller is:
CA 022~2428 1998-10-20 W O 98/00763 PCT~US97/09473 e(t) - e(t - ~t) u(t) = Kce(t) + Kd At Tuner outputs are proportional gain Kc and derivative gain Kd. The tuner input 5 consists of simple features computed from the process output after exciting the process at time to with a pulse of height H and width W. The process output is measured at times to+n 4 , for n = O, 1, . . ., 10. Ten features are computed simply as y(to+n 4 ) -y(to) for n = 1, 2,...7 10. Neither Tp nor Kp are assumed to be known in thisembodiment. However, appropriate values for H and W do depend on the process, and 10 some general notions of the time constant and gain are needed.
In all the above embodiments, the key piece of the tuner is a nonlinear approximator. Various nonlinear approximators can be used in this context, including multilayer perceptron neural networks, computationally simple sigmoidal compositions, radial basis function networks, functional link networks, CMAC networks, fuzzy logic 15 models that employ fu~ification and/or defuzzification and/or membership functions, wavelet networks, polynomial expansions, and specific nonlinear parametrized structures.
Claims (13)
1. An autotuner (21) for tuning a controller (14), comprising:
preprocessing means, having an input and an output, for transforming a set of input parameters at the input into a set of normalized parameters at the output, nonlinear approximator means (28), having an input connected to the output of said preprocessing means, for algorithmically operating on the set of normalized parameters, whereby said nonlinear approximator means is setup offline without advance supervised learning, the said setup using a system comprising a plurality of closed-loop simulations and an optimization algorithm (45), whereby during said setup the optimization algorithm minimizes the value of a cost function (44) which is derived from the said plurality of closed-loop simulations and which indicates performance of the said simulations, to produce optimized parameters of the said nonlinear approximator means, whereby the output of the said nonlinear approximator is a set of normalized tuning parameters derived on the basis of the said minimized value of said cost function; and postprocessing means, having a first input connected to the output of said nonlinear approximator means, having a second input connected to the input of the preprocessing means and having an output, for scaling the set of normalized tuning parameters into controller tuning parameters (22) at the output.
preprocessing means, having an input and an output, for transforming a set of input parameters at the input into a set of normalized parameters at the output, nonlinear approximator means (28), having an input connected to the output of said preprocessing means, for algorithmically operating on the set of normalized parameters, whereby said nonlinear approximator means is setup offline without advance supervised learning, the said setup using a system comprising a plurality of closed-loop simulations and an optimization algorithm (45), whereby during said setup the optimization algorithm minimizes the value of a cost function (44) which is derived from the said plurality of closed-loop simulations and which indicates performance of the said simulations, to produce optimized parameters of the said nonlinear approximator means, whereby the output of the said nonlinear approximator is a set of normalized tuning parameters derived on the basis of the said minimized value of said cost function; and postprocessing means, having a first input connected to the output of said nonlinear approximator means, having a second input connected to the input of the preprocessing means and having an output, for scaling the set of normalized tuning parameters into controller tuning parameters (22) at the output.
2. The tuner of claim 1 wherein:
the input parameters are either or both of two types of parameters, wherein the two types of parameters are:
parameters (26) characterizing a system to be controlled; and parameters (27) characterizing a desired closed-loop behavior;
and said nonlinear approximator means is optimized offline without supervised learning and wherein the offline optimization does not require generation of optimized timing parameters for optimizing said nonlinear approximator means.
the input parameters are either or both of two types of parameters, wherein the two types of parameters are:
parameters (26) characterizing a system to be controlled; and parameters (27) characterizing a desired closed-loop behavior;
and said nonlinear approximator means is optimized offline without supervised learning and wherein the offline optimization does not require generation of optimized timing parameters for optimizing said nonlinear approximator means.
3. The tuner of claim 2 wherein said nonlinear approximator means is a neural network.
4. The tuner of claim 2 wherein said nonlinear approximator means is a fuzzy approximator.
5. The tuner of claim 2 wherein said nonlinear approximator means is a radial basis function approximator.
6. The tuner of claim 2 wherein said nonlinear approximator means is an approximator that utilizes nontranscendental mathematical functions.
7. The tuner of claim 2 wherein said nonlinear approximator means is an approximator that utilizes computationally simple compositional sigmoidal mappings.
8. An autotuner (21) for tuning a controller (14) comprising:
a nonlinear approximator (28), the behavior of which is modified by adjusting its parameters, and which is setup off-line without advance supervised learning, the said setup using a system comprising a closed-loop control simulator and an optimization algorithm (45), whereby during said setup the closed-loop control simulator receives controller parameters from the controller for use in a plurality of closed-loop simulations, the said closed-loop simulations enabling the computation of the value of a cost function (44) which indicates performance of the closed loop simulations, whereby said value of said cost function is processed by the optimization algorithm which produces optimized nonlinear approximator parameters which are optimized in such a way as to minimize the value of said cost function, whereby said minimized value of said cost function is used to derive appropriate controller tuning parameters (22), for supply to the controller, which result in optimized controller performance.
a nonlinear approximator (28), the behavior of which is modified by adjusting its parameters, and which is setup off-line without advance supervised learning, the said setup using a system comprising a closed-loop control simulator and an optimization algorithm (45), whereby during said setup the closed-loop control simulator receives controller parameters from the controller for use in a plurality of closed-loop simulations, the said closed-loop simulations enabling the computation of the value of a cost function (44) which indicates performance of the closed loop simulations, whereby said value of said cost function is processed by the optimization algorithm which produces optimized nonlinear approximator parameters which are optimized in such a way as to minimize the value of said cost function, whereby said minimized value of said cost function is used to derive appropriate controller tuning parameters (22), for supply to the controller, which result in optimized controller performance.
9. The tuner of claim 8 wherein the nonlinear approximator is a neural network.
10. The tuner of claim 8 wherein the nonlinear approximator is a fuzzy approximator.
11. The tuner of claim 8 wherein the nonlinear approximator is a radial basis function approximator.
12. The tuner of claim 8 wherein the nonlinear approximator is an approximator that utilizes nontranscendental mathematical functions.
13. The tuner of claim 8 wherein the nonlinear approximator is an approximator that utilizes computationally simple compositional sigmoidal mappings.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US08/671,996 | 1996-06-28 | ||
| US08/671,996 US5847952A (en) | 1996-06-28 | 1996-06-28 | Nonlinear-approximator-based automatic tuner |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CA2252428A1 true CA2252428A1 (en) | 1998-01-08 |
Family
ID=24696725
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CA002252428A Abandoned CA2252428A1 (en) | 1996-06-28 | 1997-06-02 | Nonlinear-approximator-based automatic tuner |
Country Status (6)
| Country | Link |
|---|---|
| US (1) | US5847952A (en) |
| EP (1) | EP0907909B1 (en) |
| JP (1) | JP2000514217A (en) |
| CA (1) | CA2252428A1 (en) |
| DE (1) | DE69701878T2 (en) |
| WO (1) | WO1998000763A1 (en) |
Families Citing this family (87)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE19603091C1 (en) * | 1996-01-29 | 1997-07-31 | Siemens Ag | Method for controlling a controlled system, in particular an internal combustion engine |
| DE19722431A1 (en) * | 1997-05-28 | 1998-12-03 | Siemens Ag | Process for controlling a delayed process with compensation and control device for carrying out the process |
| DE19734208A1 (en) * | 1997-08-07 | 1999-02-11 | Heidenhain Gmbh Dr Johannes | Method and circuit arrangement for determining optimal controller parameters for speed control |
| US6173240B1 (en) * | 1998-11-02 | 2001-01-09 | Ise Integrated Systems Engineering Ag | Multidimensional uncertainty analysis |
| US6631299B1 (en) * | 1998-12-22 | 2003-10-07 | Texas Instruments Incorporated | System and method for self-tuning feedback control of a system |
| SG82592A1 (en) * | 1998-12-30 | 2001-08-21 | Univ Singapore | A novel predictive and self-tuning pi control apparatus for expanded process control applications |
| US8044793B2 (en) | 2001-03-01 | 2011-10-25 | Fisher-Rosemount Systems, Inc. | Integrated device alerts in a process control system |
| US6549815B1 (en) * | 1999-03-02 | 2003-04-15 | Yamaha Hatsudoki Kabushiki Kaisha | Method and apparatus for optimizing overall characteristics of device, using heuristic method |
| US6445962B1 (en) * | 1999-03-15 | 2002-09-03 | Fisher Rosemount Systems, Inc. | Auto-tuning in a distributed process control environment |
| US7085688B1 (en) * | 1999-10-22 | 2006-08-01 | Shizuo Sumida | Non-linear characteristic reproducing apparatus and non-linear characteristic reproducing program storage medium |
| US6654649B2 (en) * | 1999-12-22 | 2003-11-25 | Aspen Technology, Inc. | Computer method and apparatus for optimized controller in a non-linear process |
| US6775646B1 (en) * | 2000-02-23 | 2004-08-10 | Agilent Technologies, Inc. | Excitation signal and radial basis function methods for use in extraction of nonlinear black-box behavioral models |
| DE60102242T2 (en) * | 2000-06-29 | 2005-01-27 | Aspen Technology, Inc., Cambridge | COMPUTER PROCESS AND DEVICE FOR RESTRICTING A NON-LINEAR APPROACH APPROACH OF AN EMPIRICAL PROCESS |
| WO2002031764A2 (en) * | 2000-10-13 | 2002-04-18 | Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V. | A method for supervised teaching of a recurrent artificial neural network |
| US6954713B2 (en) * | 2001-03-01 | 2005-10-11 | Fisher-Rosemount Systems, Inc. | Cavitation detection in a process plant |
| US6795798B2 (en) | 2001-03-01 | 2004-09-21 | Fisher-Rosemount Systems, Inc. | Remote analysis of process control plant data |
| US7720727B2 (en) | 2001-03-01 | 2010-05-18 | Fisher-Rosemount Systems, Inc. | Economic calculations in process control system |
| JP4160399B2 (en) | 2001-03-01 | 2008-10-01 | フィッシャー−ローズマウント システムズ， インコーポレイテッド | Creating and displaying indicators in the process plant |
| US8073967B2 (en) | 2002-04-15 | 2011-12-06 | Fisher-Rosemount Systems, Inc. | Web services-based communications for use with process control systems |
| US7389204B2 (en) | 2001-03-01 | 2008-06-17 | Fisher-Rosemount Systems, Inc. | Data presentation system for abnormal situation prevention in a process plant |
| US7162534B2 (en) | 2001-07-10 | 2007-01-09 | Fisher-Rosemount Systems, Inc. | Transactional data communications for process control systems |
| US7117045B2 (en) * | 2001-09-08 | 2006-10-03 | Colorado State University Research Foundation | Combined proportional plus integral (PI) and neural network (nN) controller |
| DE10148097A1 (en) * | 2001-09-28 | 2003-04-17 | Duerr Systems Gmbh | Method and device for improving the response behavior of a command-controlled system |
| US20030144747A1 (en) * | 2001-11-21 | 2003-07-31 | Metso Paper Automation Oy | Method and controller to control a process |
| US7035695B2 (en) * | 2002-01-22 | 2006-04-25 | Imb Controls Inc. | Method and apparatus for tuning a PID controller |
| DE10213533A1 (en) * | 2002-03-26 | 2003-10-16 | Siemens Ag | Method and controller for regulating at least one component of a technical system |
| JP3700059B2 (en) * | 2002-08-12 | 2005-09-28 | トヨタ自動車株式会社 | Voltage conversion device, voltage conversion method, and computer-readable recording medium storing a program for causing a computer to control voltage conversion |
| US20050187643A1 (en) * | 2004-02-19 | 2005-08-25 | Pavilion Technologies, Inc. | Parametric universal nonlinear dynamics approximator and use |
| US7600234B2 (en) | 2002-12-10 | 2009-10-06 | Fisher-Rosemount Systems, Inc. | Method for launching applications |
| US8935298B2 (en) | 2002-12-30 | 2015-01-13 | Fisher-Rosemount Systems, Inc. | Integrated navigational tree importation and generation in a process plant |
| US7493310B2 (en) | 2002-12-30 | 2009-02-17 | Fisher-Rosemount Systems, Inc. | Data visualization within an integrated asset data system for a process plant |
| US6685143B1 (en) * | 2003-01-03 | 2004-02-03 | Orbital Research Inc. | Aircraft and missile forebody flow control device and method of controlling flow |
| US7152072B2 (en) | 2003-01-08 | 2006-12-19 | Fisher-Rosemount Systems Inc. | Methods and apparatus for importing device data into a database system used in a process plant |
| KR100507835B1 (en) * | 2003-02-03 | 2005-08-17 | 한국과학기술원 | A Being Most Suitable Gain Selection Method of PID Controller |
| ES2301968T3 (en) * | 2003-02-14 | 2008-07-01 | Dresser, Inc. | METHOD, SYSTEM AND MEMORY MEANS TO EFFECT THE DIAGNOSIS OF A VALVE IN A PRODUCTION LINE. |
| US7953842B2 (en) | 2003-02-19 | 2011-05-31 | Fisher-Rosemount Systems, Inc. | Open network-based data acquisition, aggregation and optimization for use with process control systems |
| US7496041B2 (en) * | 2003-02-28 | 2009-02-24 | Fisher-Rosemount Systems, Inc. | High speed auto-tuning loop |
| US7103427B2 (en) | 2003-02-28 | 2006-09-05 | Fisher-Rosemont Systems, Inc. | Delivery of process plant notifications |
| US7634384B2 (en) | 2003-03-18 | 2009-12-15 | Fisher-Rosemount Systems, Inc. | Asset optimization reporting in a process plant |
| GB2400923B (en) * | 2003-04-25 | 2005-06-01 | Falmer Investment Ltd | Adaptive fuzzy logic temperature control |
| US7299415B2 (en) | 2003-06-16 | 2007-11-20 | Fisher-Rosemount Systems, Inc. | Method and apparatus for providing help information in multiple formats |
| JP4223894B2 (en) * | 2003-08-21 | 2009-02-12 | 株式会社山武 | PID parameter adjustment device |
| JP4310804B2 (en) * | 2004-01-14 | 2009-08-12 | 横河電機株式会社 | Controller |
| US7030747B2 (en) | 2004-02-26 | 2006-04-18 | Fisher-Rosemount Systems, Inc. | Method and system for integrated alarms in a process control system |
| US7079984B2 (en) | 2004-03-03 | 2006-07-18 | Fisher-Rosemount Systems, Inc. | Abnormal situation prevention in a process plant |
| US7676287B2 (en) | 2004-03-03 | 2010-03-09 | Fisher-Rosemount Systems, Inc. | Configuration system and method for abnormal situation prevention in a process plant |
| US7515977B2 (en) | 2004-03-30 | 2009-04-07 | Fisher-Rosemount Systems, Inc. | Integrated configuration system for use in a process plant |
| US7536274B2 (en) | 2004-05-28 | 2009-05-19 | Fisher-Rosemount Systems, Inc. | System and method for detecting an abnormal situation associated with a heater |
| WO2005124491A1 (en) | 2004-06-12 | 2005-12-29 | Fisher-Rosemount Systems, Inc. | System and method for detecting an abnormal situation associated with a process gain of a control loop |
| US7181654B2 (en) | 2004-09-17 | 2007-02-20 | Fisher-Rosemount Systems, Inc. | System and method for detecting an abnormal situation associated with a reactor |
| US9201420B2 (en) | 2005-04-08 | 2015-12-01 | Rosemount, Inc. | Method and apparatus for performing a function in a process plant using monitoring data with criticality evaluation data |
| US8005647B2 (en) | 2005-04-08 | 2011-08-23 | Rosemount, Inc. | Method and apparatus for monitoring and performing corrective measures in a process plant using monitoring data with corrective measures data |
| US7283894B2 (en) | 2006-02-10 | 2007-10-16 | Dresser, Inc. | System and method for fluid regulation |
| US7912676B2 (en) | 2006-07-25 | 2011-03-22 | Fisher-Rosemount Systems, Inc. | Method and system for detecting abnormal operation in a process plant |
| US8145358B2 (en) | 2006-07-25 | 2012-03-27 | Fisher-Rosemount Systems, Inc. | Method and system for detecting abnormal operation of a level regulatory control loop |
| US8606544B2 (en) | 2006-07-25 | 2013-12-10 | Fisher-Rosemount Systems, Inc. | Methods and systems for detecting deviation of a process variable from expected values |
| US7657399B2 (en) | 2006-07-25 | 2010-02-02 | Fisher-Rosemount Systems, Inc. | Methods and systems for detecting deviation of a process variable from expected values |
| CN102789226B (en) | 2006-09-28 | 2015-07-01 | 费舍-柔斯芒特系统股份有限公司 | Abnormal situation prevention in a heat exchanger |
| US8014880B2 (en) | 2006-09-29 | 2011-09-06 | Fisher-Rosemount Systems, Inc. | On-line multivariate analysis in a distributed process control system |
| US8032341B2 (en) | 2007-01-04 | 2011-10-04 | Fisher-Rosemount Systems, Inc. | Modeling a process using a composite model comprising a plurality of regression models |
| US8032340B2 (en) | 2007-01-04 | 2011-10-04 | Fisher-Rosemount Systems, Inc. | Method and system for modeling a process variable in a process plant |
| US7539560B2 (en) * | 2007-01-05 | 2009-05-26 | Dresser, Inc. | Control valve and positioner diagnostics |
| US7827006B2 (en) | 2007-01-31 | 2010-11-02 | Fisher-Rosemount Systems, Inc. | Heat exchanger fouling detection |
| US10410145B2 (en) | 2007-05-15 | 2019-09-10 | Fisher-Rosemount Systems, Inc. | Automatic maintenance estimation in a plant environment |
| US8301676B2 (en) | 2007-08-23 | 2012-10-30 | Fisher-Rosemount Systems, Inc. | Field device with capability of calculating digital filter coefficients |
| US7702401B2 (en) | 2007-09-05 | 2010-04-20 | Fisher-Rosemount Systems, Inc. | System for preserving and displaying process control data associated with an abnormal situation |
| US9323247B2 (en) | 2007-09-14 | 2016-04-26 | Fisher-Rosemount Systems, Inc. | Personalized plant asset data representation and search system |
| US8055479B2 (en) | 2007-10-10 | 2011-11-08 | Fisher-Rosemount Systems, Inc. | Simplified algorithm for abnormal situation prevention in load following applications including plugged line diagnostics in a dynamic process |
| DE102009002682B4 (en) * | 2009-04-28 | 2012-09-20 | Airbus Operations Gmbh | Device and method for residual evaluation of a residual for detecting system errors in the system behavior of a system of an aircraft |
| CN101655688B (en) * | 2009-07-21 | 2012-03-28 | 中南大学 | Three-domain fuzzy PID control method applied in distributed parameter system |
| US9927788B2 (en) | 2011-05-19 | 2018-03-27 | Fisher-Rosemount Systems, Inc. | Software lockout coordination between a process control system and an asset management system |
| US9529348B2 (en) | 2012-01-24 | 2016-12-27 | Emerson Process Management Power & Water Solutions, Inc. | Method and apparatus for deploying industrial plant simulators using cloud computing technologies |
| TWI454868B (en) * | 2012-11-02 | 2014-10-01 | Ind Tech Res Inst | Goal-oriented computer numerical controlled automatic tuning system and method |
| KR20150050194A (en) * | 2013-10-31 | 2015-05-08 | 삼성전자주식회사 | Method of operating oscillator including verification of operation of the oscillator |
| CN103901839B (en) * | 2014-03-04 | 2018-02-09 | 清华大学 | The communication means and system of unmanned plane |
| US20150268645A1 (en) * | 2014-03-18 | 2015-09-24 | Honeywell Asca Inc. | Method and apparatus for specifying and visualizing robust tuning of model-based controllers |
| CN104714408B (en) * | 2014-12-31 | 2017-04-05 | 重庆科技学院 | The acquisition methods and device of lag-lead-corrector structural parameters |
| CN104635725B (en) * | 2014-12-31 | 2017-05-17 | 重庆科技学院 | Method and device for obtaining lead corrector structure parameters |
| US10542961B2 (en) | 2015-06-15 | 2020-01-28 | The Research Foundation For The State University Of New York | System and method for infrasonic cardiac monitoring |
| US10659364B2 (en) * | 2017-04-19 | 2020-05-19 | Nxp Usa, Inc. | Hybrid AQM controller using RBF for queue utilization approximation as a non-linear system |
| CN107290958B (en) * | 2017-04-24 | 2022-08-26 | 南京工业大学 | Novel slewing bearing test bed drive control method |
| US11879656B2 (en) * | 2018-04-04 | 2024-01-23 | International Business Machines Corporation | Initialization of radial base function neural network nodes for reinforcement learning incremental control system |
| CN109656138A (en) * | 2018-12-19 | 2019-04-19 | 龙口盛福达食品有限公司 | A kind of brewed spirit process temperature intelligent control method based on fuzzy reasoning |
| CN110908667B (en) * | 2019-11-18 | 2021-11-16 | 北京迈格威科技有限公司 | Method and device for joint compilation of neural network and electronic equipment |
| EP3825788B1 (en) * | 2019-11-19 | 2022-11-09 | Asco Numatics GmbH | Control device, control system and control method for regulating a physical quantity of a fluid |
| KR102318164B1 (en) | 2020-01-20 | 2021-10-26 | 두산중공업 주식회사 | Apparatus and method for automatically tuning a fluid temperature PID controller with constraints on the physical characteristics of the process |
| GB2620635A (en) * | 2022-07-15 | 2024-01-17 | Ucl Business Ltd | PID tuning |
Family Cites Families (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| AU560995B2 (en) * | 1984-02-07 | 1987-04-30 | Toshiba, Kabushiki Kaisha | Process control apparatus |
| US4677466A (en) * | 1985-07-29 | 1987-06-30 | A. C. Nielsen Company | Broadcast program identification method and apparatus |
| DE3811086A1 (en) * | 1987-04-03 | 1988-10-20 | Hitachi Ltd | PID CONTROL SYSTEM |
| EP0360206A3 (en) * | 1988-09-21 | 1990-12-19 | Hitachi, Ltd. | Self-tuning controller apparatus and process control system |
| CA2031765C (en) * | 1989-12-08 | 1996-02-20 | Masahide Nomura | Method and system for performing control conforming with characteristics of controlled system |
| US5111531A (en) * | 1990-01-08 | 1992-05-05 | Automation Technology, Inc. | Process control using neural network |
| US5295061A (en) * | 1990-04-20 | 1994-03-15 | Sanyo Electric Co., Ltd. | Control parameter tuning unit and a method of tuning parameters for a control unit |
| EP0460892B1 (en) * | 1990-06-04 | 1996-09-04 | Hitachi, Ltd. | A control device for controlling a controlled apparatus, and a control method therefor |
| US5159660A (en) * | 1990-08-09 | 1992-10-27 | Western Thunder | Universal process control using artificial neural networks |
| US5175678A (en) * | 1990-08-15 | 1992-12-29 | Elsag International B.V. | Method and procedure for neural control of dynamic processes |
| US5159547A (en) * | 1990-10-16 | 1992-10-27 | Rockwell International Corporation | Self-monitoring tuner for feedback controller |
| US5283729A (en) * | 1991-08-30 | 1994-02-01 | Fisher-Rosemount Systems, Inc. | Tuning arrangement for turning the control parameters of a controller |
| US5513098A (en) * | 1993-06-04 | 1996-04-30 | The Johns Hopkins University | Method for model-free control of general discrete-time systems |
| DE4433332A1 (en) * | 1994-09-19 | 1996-03-21 | Siemens Ag | Device for adaptive control of a route |
-
1996
- 1996-06-28 US US08/671,996 patent/US5847952A/en not_active Expired - Lifetime
-
1997
- 1997-06-02 CA CA002252428A patent/CA2252428A1/en not_active Abandoned
- 1997-06-02 DE DE69701878T patent/DE69701878T2/en not_active Expired - Fee Related
- 1997-06-02 JP JP10504126A patent/JP2000514217A/en active Pending
- 1997-06-02 WO PCT/US1997/009473 patent/WO1998000763A1/en active IP Right Grant
- 1997-06-02 EP EP97929758A patent/EP0907909B1/en not_active Expired - Lifetime
Also Published As
| Publication number | Publication date |
|---|---|
| DE69701878T2 (en) | 2000-12-07 |
| US5847952A (en) | 1998-12-08 |
| JP2000514217A (en) | 2000-10-24 |
| EP0907909B1 (en) | 2000-05-03 |
| DE69701878D1 (en) | 2000-06-08 |
| EP0907909A1 (en) | 1999-04-14 |
| WO1998000763A1 (en) | 1998-01-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CA2252428A1 (en) | Nonlinear-approximator-based automatic tuner | |
| Zhang et al. | Design and Implementation of hybrid modeling and PFC for oxygen content regulation in a coke furnace | |
| Perrusquia et al. | Discrete-time H 2 neural control using reinforcement learning | |
| Zhang et al. | Adaptive optimal dynamic surface control of strict‐feedback nonlinear systems with output constraints | |
| Korupu et al. | A comparative study of various Smith predictor configurations for industrial delay processes | |
| Zhang et al. | Adaptive neural control of unknown non-affine nonlinear systems with input deadzone and unknown disturbance | |
| Chakraborty et al. | Inverse plant model and frequency loop shaping-based PID controller design for processes with time-delay | |
| Xu et al. | Terminal sliding mode control using adaptive fuzzy-neural observer | |
| Sousa et al. | Level control of coupled tank system based on neural network techniques | |
| Mohammadzadeh et al. | Modern Adaptive Fuzzy Control Systems | |
| Revelo et al. | Nonsquare multivariable chemical processes: a hybrid centralized control proposal | |
| Haber et al. | Controlling a complex electromechanical process on the basis of a neurofuzzy approach | |
| Mahapatro et al. | Design and real‐time implementation of an adaptive fuzzy sliding mode controller for a coupled tank system | |
| Sivadasan et al. | Design of cross-coupled nonlinear PID controller using single-objective evolutionary algorithms | |
| Ghadiri et al. | Predictive functional control based on particle swarm optimisation algorithm for MIMO process with time delay | |
| Yan et al. | Model based control using artificial neural networks | |
| Yazdanpanah et al. | Design PID baseline fuzzy tuning proportional-derivative coefficient nonlinear controller with application to continuum robot | |
| Santhosh et al. | Fractional-Order Controller design for Electrical Systems-A Study | |
| Tarantino et al. | Finite-horizon optimal control for linear and nonlinear systems relying on constant optimal costate | |
| Xu et al. | Cooperative Control of Recurrent Neural Network for PID-based Single Phase Hotplate Temperature Control Systems | |
| Gupta et al. | Machine learning approach based on fuzzy logic for industrial temperature regulation | |
| Said et al. | ARX Modeling and FPID with model reference controller design of a hot air blower system | |
| Pappa et al. | Real-time implementation of model predictive neural controller for heat exchanger | |
| Chew et al. | Evolutional Based Optimization Analysis for Three-element Control System | |
| Souto et al. | Use of Feedback Loop Control Theory in Software Project Productivity Control A Simulation Study. |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| EEER | Examination request | ||
| FZDE | Discontinued |